bpo-17611. Move unwinding of stack for "pseudo exceptions" from inter…

…preter to compiler. (GH-5006)



Co-authored-by: Mark Shannon <mark@hotpy.org>
Co-authored-by: Antoine Pitrou <antoine@python.org>

Doc/library/dis.rst

@@ -335,6 +335,14 @@ The Python compiler currently generates the following bytecode instructions.
    three.
 
 
+.. opcode:: ROT_FOUR
+
+   Lifts second, third and forth stack items one position up, moves top down
+   to position four.
+
+   .. versionadded:: 3.8
+
+
 .. opcode:: DUP_TOP
 
    Duplicates the reference on top of the stack.
@@ -605,17 +613,6 @@ the original TOS1.
    is terminated with :opcode:`POP_TOP`.
 
 
-.. opcode:: BREAK_LOOP
-
-   Terminates a loop due to a :keyword:`break` statement.
-
-
-.. opcode:: CONTINUE_LOOP (target)
-
-   Continues a loop due to a :keyword:`continue` statement.  *target* is the
-   address to jump to (which should be a :opcode:`FOR_ITER` instruction).
-
-
 .. opcode:: SET_ADD (i)
 
    Calls ``set.add(TOS1[-i], TOS)``.  Used to implement set comprehensions.
@@ -676,7 +673,7 @@ iterations of the loop.
 .. opcode:: POP_BLOCK
 
    Removes one block from the block stack.  Per frame, there is a stack of
-   blocks, denoting nested loops, try statements, and such.
+   blocks, denoting :keyword:`try` statements, and such.
 
 
 .. opcode:: POP_EXCEPT
@@ -687,11 +684,50 @@ iterations of the loop.
    popped values are used to restore the exception state.
 
 
+.. opcode:: POP_FINALLY (preserve_tos)
+
+   Cleans up the value stack and the block stack.  If *preserve_tos* is not
+   ``0`` TOS first is popped from the stack and pushed on the stack after
+   perfoming other stack operations:
+
+   * If TOS is ``NULL`` or an integer (pushed by :opcode:`BEGIN_FINALLY`
+     or :opcode:`CALL_FINALLY`) it is popped from the stack.
+   * If TOS is an exception type (pushed when an exception has been raised)
+     6 values are popped from the stack, the last three popped values are
+     used to restore the exception state.  An exception handler block is
+     removed from the block stack.
+
+   It is similar to :opcode:`END_FINALLY`, but doesn't change the bytecode
+   counter nor raise an exception.  Used for implementing :keyword:`break`
+   and :keyword:`return` in the :keyword:`finally` block.
+
+   .. versionadded:: 3.8
+
+
+.. opcode:: BEGIN_FINALLY
+
+   Pushes ``NULL`` onto the stack for using it in :opcode:`END_FINALLY`,
+   :opcode:`POP_FINALLY`, :opcode:`WITH_CLEANUP_START` and
+   :opcode:`WITH_CLEANUP_FINISH`.  Starts the :keyword:`finally` block.
+
+   .. versionadded:: 3.8
+
+
 .. opcode:: END_FINALLY
 
    Terminates a :keyword:`finally` clause.  The interpreter recalls whether the
-   exception has to be re-raised, or whether the function returns, and continues
-   with the outer-next block.
+   exception has to be re-raised or execution has to be continued depending on
+   the value of TOS.
+
+   * If TOS is ``NULL`` (pushed by :opcode:`BEGIN_FINALLY`) continue from
+     the next instruction. TOS is popped.
+   * If TOS is an integer (pushed by :opcode:`CALL_FINALLY`), sets the
+     bytecode counter to TOS.  TOS is popped.
+   * If TOS is an exception type (pushed when an exception has been raised)
+     6 values are popped from the stack, the first three popped values are
+     used to re-raise the exception and the last three popped values are used
+     to restore the exception state.  An exception handler block is removed
+     from the block stack.
 
 
 .. opcode:: LOAD_BUILD_CLASS
@@ -704,9 +740,9 @@ iterations of the loop.
 
    This opcode performs several operations before a with block starts.  First,
    it loads :meth:`~object.__exit__` from the context manager and pushes it onto
-   the stack for later use by :opcode:`WITH_CLEANUP`.  Then,
+   the stack for later use by :opcode:`WITH_CLEANUP_START`.  Then,
    :meth:`~object.__enter__` is called, and a finally block pointing to *delta*
-   is pushed.  Finally, the result of calling the enter method is pushed onto
+   is pushed.  Finally, the result of calling the ``__enter__()`` method is pushed onto
    the stack.  The next opcode will either ignore it (:opcode:`POP_TOP`), or
    store it in (a) variable(s) (:opcode:`STORE_FAST`, :opcode:`STORE_NAME`, or
    :opcode:`UNPACK_SEQUENCE`).
@@ -716,30 +752,31 @@ iterations of the loop.
 
 .. opcode:: WITH_CLEANUP_START
 
-   Cleans up the stack when a :keyword:`with` statement block exits.  TOS is the
-   context manager's :meth:`__exit__` bound method. Below TOS are 1--3 values
-   indicating how/why the finally clause was entered:
+   Starts cleaning up the stack when a :keyword:`with` statement block exits.
 
-   * SECOND = ``None``
-   * (SECOND, THIRD) = (``WHY_{RETURN,CONTINUE}``), retval
-   * SECOND = ``WHY_*``; no retval below it
-   * (SECOND, THIRD, FOURTH) = exc_info()
+   At the top of the stack are either ``NULL`` (pushed by
+   :opcode:`BEGIN_FINALLY`) or 6 values pushed if an exception has been
+   raised in the with block.  Below is the context manager's
+   :meth:`~object.__exit__` or :meth:`~object.__aexit__` bound method.
 
-   In the last case, ``TOS(SECOND, THIRD, FOURTH)`` is called, otherwise
-   ``TOS(None, None, None)``.  Pushes SECOND and result of the call
-   to the stack.
+   If TOS is ``NULL``, calls ``SECOND(None, None, None)``,
+   removes the function from the stack, leaving TOS, and pushes ``None``
+   to the stack.  Otherwise calls ``SEVENTH(TOP, SECOND, THIRD)``,
+   shifts the bottom 3 values of the stack down, replaces the empty spot
+   with ``NULL`` and pushes TOS.  Finally pushes the result of the call.
 
 
 .. opcode:: WITH_CLEANUP_FINISH
 
-   Pops exception type and result of 'exit' function call from the stack.
+   Finishes cleaning up the stack when a :keyword:`with` statement block exits.
 
-   If the stack represents an exception, *and* the function call returns a
-   'true' value, this information is "zapped" and replaced with a single
-   ``WHY_SILENCED`` to prevent :opcode:`END_FINALLY` from re-raising the
-   exception.  (But non-local gotos will still be resumed.)
+   TOS is result of ``__exit__()`` or ``__aexit__()`` function call pushed
+   by :opcode:`WITH_CLEANUP_START`.  SECOND is ``None`` or an exception type
+   (pushed when an exception has been raised).
 
-   .. XXX explain the WHY stuff!
+   Pops two values from the stack.  If SECOND is not None and TOS is true
+   unwinds the EXCEPT_HANDLER block which was created when the exception
+   was caught and pushes ``NULL`` to the stack.
 
 
 All of the following opcodes use their arguments.
@@ -987,22 +1024,19 @@ All of the following opcodes use their arguments.
    Loads the global named ``co_names[namei]`` onto the stack.
 
 
-.. opcode:: SETUP_LOOP (delta)
-
-   Pushes a block for a loop onto the block stack.  The block spans from the
-   current instruction with a size of *delta* bytes.
-
+.. opcode:: SETUP_FINALLY (delta)
 
-.. opcode:: SETUP_EXCEPT (delta)
+   Pushes a try block from a try-finally or try-except clause onto the block
+   stack.  *delta* points to the finally block or the first except block.
 
-   Pushes a try block from a try-except clause onto the block stack. *delta*
-   points to the first except block.
 
+.. opcode:: CALL_FINALLY (delta)
 
-.. opcode:: SETUP_FINALLY (delta)
+   Pushes the address of the next instruction onto the stack and increments
+   bytecode counter by *delta*.  Used for calling the finally block as a
+   "subroutine".
 
-   Pushes a try block from a try-except clause onto the block stack. *delta*
-   points to the finally block.
+   .. versionadded:: 3.8
 
 
 .. opcode:: LOAD_FAST (var_num)

Doc/whatsnew/3.8.rst

@@ -137,3 +137,21 @@ Changes in the Python API
   :func:`dbm.dumb.open` with flags ``'r'`` and ``'w'`` no longer creates
   a database if it does not exist.
   (Contributed by Serhiy Storchaka in :issue:`32749`.)
+
+
+CPython bytecode changes
+------------------------
+
+* The interpreter loop  has been simplified by moving the logic of unrolling
+  the stack of blocks into the compiler.  The compiler emits now explicit
+  instructions for adjusting the stack of values and calling the cleaning
+  up code for :keyword:`break`, :keyword:`continue` and :keyword:`return`.
+
+  Removed opcodes :opcode:`BREAK_LOOP`, :opcode:`CONTINUE_LOOP`,
+  :opcode:`SETUP_LOOP` and :opcode:`SETUP_EXCEPT`.  Added new opcodes
+  :opcode:`ROT_FOUR`, :opcode:`BEGIN_FINALLY`, :opcode:`CALL_FINALLY` and
+  :opcode:`POP_FINALLY`.  Changed the behavior of :opcode:`END_FINALLY`
+  and :opcode:`WITH_CLEANUP_START`.
+
+  (Contributed by Mark Shannon, Antoine Pitrou and Serhiy Storchaka in
+  :issue:`17611`.)

Include/opcode.h

@@ -12,6 +12,7 @@ extern "C" {
 #define ROT_THREE                 3
 #define DUP_TOP                   4
 #define DUP_TOP_TWO               5
+#define ROT_FOUR                  6
 #define NOP                       9
 #define UNARY_POSITIVE           10
 #define UNARY_NEGATIVE           11
@@ -32,6 +33,7 @@ extern "C" {
 #define GET_AITER                50
 #define GET_ANEXT                51
 #define BEFORE_ASYNC_WITH        52
+#define BEGIN_FINALLY            53
 #define INPLACE_ADD              55
 #define INPLACE_SUBTRACT         56
 #define INPLACE_MULTIPLY         57
@@ -55,7 +57,6 @@ extern "C" {
 #define INPLACE_AND              77
 #define INPLACE_XOR              78
 #define INPLACE_OR               79
-#define BREAK_LOOP               80
 #define WITH_CLEANUP_START       81
 #define WITH_CLEANUP_FINISH      82
 #define RETURN_VALUE             83
@@ -92,9 +93,6 @@ extern "C" {
 #define POP_JUMP_IF_FALSE       114
 #define POP_JUMP_IF_TRUE        115
 #define LOAD_GLOBAL             116
-#define CONTINUE_LOOP           119
-#define SETUP_LOOP              120
-#define SETUP_EXCEPT            121
 #define SETUP_FINALLY           122
 #define LOAD_FAST               124
 #define STORE_FAST              125
@@ -127,6 +125,8 @@ extern "C" {
 #define BUILD_TUPLE_UNPACK_WITH_CALL 158
 #define LOAD_METHOD             160
 #define CALL_METHOD             161
+#define CALL_FINALLY            162
+#define POP_FINALLY             163
 
 /* EXCEPT_HANDLER is a special, implicit block type which is created when
    entering an except handler. It is not an opcode but we define it here

Lib/importlib/_bootstrap_external.py

@@ -246,6 +246,7 @@ def _write_atomic(path, data, mode=0o666):
 #     Python 3.7a2  3391 (update GET_AITER #31709)
 #     Python 3.7a4  3392 (PEP 552: Deterministic pycs #31650)
 #     Python 3.7b1  3393 (remove STORE_ANNOTATION opcode #32550)
+#     Python 3.8a1  3400 (move frame block handling to compiler #17611)
 #
 # MAGIC must change whenever the bytecode emitted by the compiler may no
 # longer be understood by older implementations of the eval loop (usually
@@ -254,7 +255,7 @@ def _write_atomic(path, data, mode=0o666):
 # Whenever MAGIC_NUMBER is changed, the ranges in the magic_values array
 # in PC/launcher.c must also be updated.
 
-MAGIC_NUMBER = (3393).to_bytes(2, 'little') + b'\r\n'
+MAGIC_NUMBER = (3400).to_bytes(2, 'little') + b'\r\n'
 _RAW_MAGIC_NUMBER = int.from_bytes(MAGIC_NUMBER, 'little')  # For import.c
 
 _PYCACHE = '__pycache__'

Lib/opcode.py

@@ -60,6 +60,7 @@ def jabs_op(name, op):
 def_op('ROT_THREE', 3)
 def_op('DUP_TOP', 4)
 def_op('DUP_TOP_TWO', 5)
+def_op('ROT_FOUR', 6)
 
 def_op('NOP', 9)
 def_op('UNARY_POSITIVE', 10)
@@ -86,6 +87,7 @@ def jabs_op(name, op):
 def_op('GET_AITER', 50)
 def_op('GET_ANEXT', 51)
 def_op('BEFORE_ASYNC_WITH', 52)
+def_op('BEGIN_FINALLY', 53)
 
 def_op('INPLACE_ADD', 55)
 def_op('INPLACE_SUBTRACT', 56)
@@ -113,10 +115,8 @@ def jabs_op(name, op):
 def_op('INPLACE_AND', 77)
 def_op('INPLACE_XOR', 78)
 def_op('INPLACE_OR', 79)
-def_op('BREAK_LOOP', 80)
 def_op('WITH_CLEANUP_START', 81)
 def_op('WITH_CLEANUP_FINISH', 82)
-
 def_op('RETURN_VALUE', 83)
 def_op('IMPORT_STAR', 84)
 def_op('SETUP_ANNOTATIONS', 85)
@@ -158,10 +158,7 @@ def jabs_op(name, op):
 
 name_op('LOAD_GLOBAL', 116)     # Index in name list
 
-jabs_op('CONTINUE_LOOP', 119)   # Target address
-jrel_op('SETUP_LOOP', 120)      # Distance to target address
-jrel_op('SETUP_EXCEPT', 121)    # ""
-jrel_op('SETUP_FINALLY', 122)   # ""
+jrel_op('SETUP_FINALLY', 122)   # Distance to target address
 
 def_op('LOAD_FAST', 124)        # Local variable number
 haslocal.append(124)
@@ -213,5 +210,7 @@ def jabs_op(name, op):
 
 name_op('LOAD_METHOD', 160)
 def_op('CALL_METHOD', 161)
+jrel_op('CALL_FINALLY', 162)
+def_op('POP_FINALLY', 163)
 
 del def_op, name_op, jrel_op, jabs_op